USE GEMMA TO TRANSLATION

#37
by tomsum - opened

May I ask if you have any valuable experience in commenting on gemma’s translations?

Hi @tomsum
Thanks for the issue, personally I did not tried gemma for translation, but my gut feeling is that with a bit of prompt engineering you'll be able to achieve interesting results. You can also try out the instruction fine-tuned model. I will let the model authors add some additional details if they have experienced anything interesting with translation using this model

Google org

Hey there -- are there any prompts or workflows that you've tried, or would like to see work better? Please let us know!

Hi,

I've tried some translation with Gemma-7b. I've used very straightforward prompt--
"translate english to assamese.
english sentence: farming is the primary occupation of man.
assamese sentence: কৃষি মানুহৰ প্ৰধান জীৱিকা।
english sentence: the human society cannot think of food without farming.
assamese sentence: কৃষি অবিহনে মানৱ সমাজে খাদ্যৰ কথা ভাবিব নোৱাৰে।
english sentence:
indigenous fruits of assam such as leteku, ponyal, kuji thekera, rupahi thekera, thereju, mirika tenga, bamboo etc., which have high edible and medicinal value.
assamese sentence:
অসমৰ থলুৱা ফল যেনে লেতেকু, পনিয়ল, কুঁজি থেকেৰা, ৰূপহী থেকেৰা, ঠেৰেজু, মিৰিকা টেঙা ইত্যাদিৰ খাদ্য আৰু ঔষধি গুণ অতি বেছি ।
english sentence: agricultural production is low in assam each time due to the severe damage caused by flood.
assamese sentence: "

Result=অসমত কৃষি উৎপাদন ক্ৰমাগতভাৱে কম হোৱাৰ কাৰণ ব্ৰহ্মপুত্ৰ আৰু আন নদীৰ বহুতো কালি ক্ষতি কৰা ।
The translation direction was English to Assamese (a low-resource Indo-Aryan Indic language). The result was not so good. But the results in reverse direction was a little bit encouraging

Google org

ah thanks for the feedback; we're thinking about ways to improve our multilingual performance. are there multilingual datasets or results that you find compelling that we should consider including for future work?

ah thanks for the feedback; we're thinking about ways to improve our multilingual performance. are there multilingual datasets or results that you find compelling that we should consider including for future work?

Hi Surya,
First of all, thank you all for the wonderful work. I'am currently doing my doctoral studies in MNMT for low resource indic languages. We are exploring ways to leverage LLMs for our translation task. For multilingual dataset, as I'm aware of in the context of NMT and Indic languages, datasets from ai4bharat is a great one. NLLB datasets are another great resource.
As per my experiments, I found gemma to be best for translation for Assamese to English direction till now. I've tried LLAMA 2 7B, BLOOM, ChatGPT 3.5 too.

Google org

that's great to hear, especially if you find multilingual competitive with chatgpt 3.5! thanks for your hard work here, and for the pointers :)

Sign up or log in to comment